Global Receptive Fields for Placental Ultrasound in HBV Vertical Transmission Risk

Hepatitis B virus (HBV) is passed from mother to child at delivery or, in a subset of pregnancies, across the placenta before birth, and this vertical route is what keeps the global reservoir of chronic infection alive: roughly half of the world's ~350 million chronic carriers acquired the virus from their mothers, and infants exposed at birth become lifelong carriers 70–90% of the time when their mothers carry the replicative e-antigen, versus the near-universal clearance seen in adults HEPATITIS B. IN PREGNANCY. The clinical defense works well — third-trimester antivirals for highly viremic mothers plus immediate newborn antibody and vaccine administration block more than 85% of transmissions — but the residual failures concentrate in exactly the mothers that the blood-marker rule already flags as high risk, and the mechanisms of in-utero passage through the placenta remain unmeasured by any current modality. Meanwhile, in high-endemicity, low-resource regions such as Bo, Sierra Leone, where 13.7% of tested individuals are surface-antigen positive and prevalence is flat across all age brackets — the signature of unmitigated historical transmission in infancy — quantitative viral-load testing is often simply unavailable Seroprevalence of hepatitis B surface antigen (HBsAg) in Bo, Sierra Leone, 2012–2013.

The one organ that sits physically between a viremic mother and her fetus is the placenta, and it can be imaged with cheap, portable ultrasound. But reading a placenta from ultrasound is precisely where a well-known architectural weakness bites. Convolutional networks, the standard image-analysis tool, compute each pixel's representation from a small neighborhood, and that neighborhood grows only linearly with depth,

$$\text{receptive field} \propto L \cdot (K - 1),$$

where $L$ is the number of stacked layers and $K$ the kernel width, so relating structures that are far apart on the image — a basal-plate margin on one edge and a retroplacental clear space on the other — forces the network through downsampling that permanently discards fine boundary detail Non-locally Enhanced Encoder-Decoder Network for Single Image De-raining. Prior obstetric work sidestepped this by physically stitching multiple probes into a wide field of view, yet still processed the stitched canvas with ordinary convolutions, and the resulting network collapsed when imaging conditions shifted (posterior-placenta Dice fell to 0.26) because local kernels latch onto acoustic artifacts rather than anatomy Placenta Segmentation in Ultrasound Imaging: Addressing Sources of Uncertainty and Limited Field-of-View. This proposal argues that the local receptive field limitation is the binding constraint for placental imaging in the vertical-transmission problem, and that removing it — with attention whose cost stays within a portable hardware budget — both fixes the known failure mode and opens the first image-informed channel for predicting which HBV-exposed infants current prophylaxis will fail.

Core Claim

The claim has one architectural half and one clinical half. Architecturally: on wide-field and out-of-distribution placental ultrasound, inserting factorized global attention into a standard encoder will close most of the domain-shift gap that pure convolutions exhibit (the 0.26-to-0.45 posterior-placenta collapse), because global context lets the model represent whole-organ geometry instead of shadowing cues, and it will do so at a cost that dense non-local blocks cannot afford. Clinically: placental image features extracted this way carry risk information about residual vertical transmission that maternal serology alone does not, so a fused model beats the blood-marker rule that current care relies on. Both halves are falsifiable, and the second is stated as a hypothesis to be tested, not assumed.

Research Program

Stage one is pure architecture work on the only public obstetric ultrasound benchmark in the literature we have: the placental segmentation datasets of 1,188 in-country images from 71 patients, 292 images from 57 patients at a second site, and a 53-image inter-operator variability set Placenta Segmentation in Ultrasound Imaging: Addressing Sources of Uncertainty and Limited Field-of-View. We keep that paper's multi-task design (placental-location classification as an auxiliary task, which already lifted out-of-distribution posterior Dice from 0.26 to 0.45) and swap the convolutional backbone's mid-stages for criss-cross attention: instead of every pixel attending to every other pixel — which costs memory and computation proportional to $N^2$ for $N$ pixels — each pixel attends only along its own row and column, and a second repetition of the same cheap operation propagates context across the entire image, giving a full-image receptive field at cost $O(N\sqrt{N})$, an 11-fold memory and 85% compute reduction versus a dense non-local block CCNet: Criss-Cross Attention for Semantic Segmentation. Two design points follow from the neighboring literature. First, we do not discard convolutions: hybrid designs that keep local inductive bias while adding global mixing — a Transformer bottleneck over a convolutional encoder, or depth-wise convolutions inside the attention network's feed-forward stage — consistently beat both pure convolutions and pure Transformers on medical images ViT-V-Net: Vision Transformer for Unsupervised Volumetric Medical Image Registration, ConvFormer: Combining CNN and Transformer for Medical Image Segmentation, and the same lesson holds for state-space models that avoid flattening 2D scans into 1D sequences EccoMamba: Enhanced Cross-hierarchical Continuity Orthogonal Mamba for Medical Image Segmentation. Second, we build a simulated wide-field test: by compositing the existing images into canvases of growing width (mimicking multi-sweep stitching of the kind the placenta paper performed in hardware), we can measure how segmentation accuracy degrades with canvas size for a local network versus the attention variant. The prediction is that the local network degrades as the canvas grows — context must survive ever more pooling — while the attention variant stays flat, which would directly demonstrate that receptive fields, not sensor physics, are the ceiling.

Stage two is the clinical study, designed around the epidemiology of the application setting. In intermediate-endemicity China, 6.64% of pregnant women are surface-antigen positive, a quarter of those are e-antigen positive, and 26.73% carry viral loads above the $2\times10^5$ IU/ml threshold that triggers third-trimester tenofovir Hepatitis B Infection Among Pregnant Women in China: A Systematic Review and Meta-Analysis. We would enroll HBsAg-positive pregnancies at antenatal screening, follow standard prophylaxis, and add point-of-care placental ultrasound sweeps at 28–36 weeks plus the routine serology panel. The primary endpoint cannot be overt infant infection: under full prophylaxis that occurred in only 1.43% of children (3 of 210) in the Jiangsu cohort Rare Detection of Occult Hepatitis B Virus Infection in Children of Mothers with Positive Hepatitis B Surface Antigen, an event rate far too low to power a modest cohort. Instead we use in-utero exposure surrogates with higher event rates — cord-blood HBV DNA and infant anti-HBc at birth — with overt infection at 6 and 12 months as the secondary endpoint, verified by pre-S/S gene sequence comparison between each mother-child pair. That verification step is not optional: in the Jiangsu study, 9 of 207 immunized children initially tested PCR-positive for viral DNA, but 7 of those 9 carried sequences identical to unrelated children — laboratory cross-contamination, not transmission — and only phylogenetic homology checking exposed this, driving the true occult-infection rate to 0% Rare Detection of Occult Hepatitis B Virus Infection in Children of Mothers with Positive Hepatitis B Surface Antigen. The fusion model concatenates the attention-extracted image embedding with the serology vector and predicts the exposure surrogate; the decisive comparison is against the same predictor with the image branch removed.

Baselines

The "do nothing" reference point is current standard of care: the threshold rule that flags mothers by e-antigen positivity or viral load above $2\times10^5$ IU/ml and escalates to tenofovir HEPATITIS B. IN PREGNANCY. As a predictor it is a hard cutoff on blood markers alone; no reported result found on Lacuna for its discrimination (AUC or similar) against immunoprophylaxis failure, so we will estimate its operating point on our own cohort, using the published anchors (>85% protection from combined immunoprophylaxis; 70–90% chronicity without it, for e-antigen-positive mothers) as the clinical context. Any learned model must beat this rule on calibration and net benefit, not just AUC, or it has no clinical role.

The strongest existing method on the imaging side is the multi-task segmentation network of the placenta paper, which is the current best published result on exactly our benchmark: in-distribution Dice of 0.86 for anterior and 0.83 for posterior placentas, and out-of-distribution posterior Dice of 0.45 with multi-task training Placenta Segmentation in Ultrasound Imaging: Addressing Sources of Uncertainty and Limited Field-of-View. We will rerun it ourselves on identical splits with an equal tuning budget and matched parameter count, because our claim is about the architecture, not their engineering. Beating it means holding in-distribution Dice at parity while lifting out-of-distribution posterior Dice materially above 0.45 — the regime where local kernels demonstrably fail.

For the risk-prediction task, the strongest published ultrasound pipeline in HBV liver disease is the ResNet-50 with clinical region-of-interest masking, view-specific normalization, and statistical pooling across views, which reached a partial AUC of 0.783 and recall-at-90%-precision of 63.4% on 610 HBV patient studies, up from 0.710 and 41.0% for the plain backbone Reliable Liver Fibrosis Assessment from Ultrasound using Global Hetero-Image Fusion and View-Specific Parameterization. Its pooling is deliberately order- and space-invariant, which makes it the ideal local-only, attention-free comparator: we adapt its fusion head onto our placental sweeps so that any gain is attributable to global receptive fields rather than to better data hygiene. The tabular side is anchored by the EHR study in which XGBoost over selected features achieved a 1-year cirrhosis AUC of 0.83, beating both a foundation Transformer pre-trained on 55 million patients (0.80) and a from-scratch Transformer (0.74) A multi-domain EHR foundation model for predicting Hepatitis B liver disease: a clinical perspective; we run the equivalent gradient-boosted model on our serology panel as the no-imaging ceiling.

Finally, the ablation that isolates our key ingredient: the identical hybrid backbone with the attention stage removed, and with one versus two criss-cross repetitions. One repetition gives each pixel only row-and-column context; only the second completes the full-image receptive field CCNet: Criss-Cross Attention for Semantic Segmentation, so a jump from one repetition to two, with no other change, is the cleanest evidence that long-range context — not added capacity — is what matters. Prior ablations support the expectation: removing the global attention component from a hybrid medical segmentation network cost 9.6% F1 on ultrasound ConvFormer: Combining CNN and Transformer for Medical Image Segmentation, and a hybrid Transformer-convolution model beat the purely convolutional VoxelMorph-2 on 3D brain registration (mean Dice 0.726 versus 0.711) precisely on large-displacement, long-range structure ViT-V-Net: Vision Transformer for Unsupervised Volumetric Medical Image Registration.

Why not just…?

Why not just use the blood markers, or XGBoost on them? This is the most serious objection, and the evidence behind it is strong: in the largest HBV modeling study available, simple gradient boosting beat every deep alternative A multi-domain EHR foundation model for predicting Hepatitis B liver disease: a clinical perspective. But that study modeled adult cirrhosis and cancer from billing codes; it says nothing about the placenta, and the clinical literature concedes that serum titers do not explain the residual in-utero transmissions that occur in highly viremic mothers despite prophylaxis HEPATITIS B. IN PREGNANCY. In settings like Bo, Sierra Leone, quantitative PCR may not exist at all, while rapid tests with 98% sensitivity and 99% specificity and portable ultrasound do Seroprevalence of hepatitis B surface antigen (HBsAg) in Bo, Sierra Leone, 2012–2013. We settle this empirically and report it either way: if the nested image-plus-serology model does not add AUC or net benefit over serology-only XGBoost on our cohort, the clinical half of the claim is dead and we will say so.

Why not just stitch more probes instead of changing the network? Hardware stitching widens the physical window but leaves the neural bottleneck untouched, and the published result shows the consequence: even on stitched wide-field images, the convolutional network's out-of-distribution posterior Dice collapsed to 0.26 Placenta Segmentation in Ultrasound Imaging: Addressing Sources of Uncertainty and Limited Field-of-View. Custom multi-probe holders also demand hardware and calibration that community clinics in endemic regions do not have, whereas attention is a software change. Our simulated wide-field experiment directly tests the skeptic's version: if a local network on wide canvases matches the attention variant, we concede the architectural half.

Why not just use a dense non-local block or a plain Vision Transformer? Dense non-local attention gives every pixel access to every other pixel in one step, but at an $N^2$ cost that the criss-cross factorization cuts by 11x in memory and 85% in compute CCNet: Criss-Cross Attention for Semantic Segmentation — a difference that decides whether the model runs on a portable machine at a district hospital. Pure Transformers additionally lose voxel-level localization through patch projection, which is why hybrids dominate in medical imaging ViT-V-Net: Vision Transformer for Unsupervised Volumetric Medical Image Registration. Still, we treat this as an open comparison rather than a settled one: a ConvFormer-style variant and a small ViT baseline run at matched parameters are part of the plan, and if one of them matches criss-cross performance within the deployment budget, we adopt it — our claim is that global context is necessary, not that one factorization is uniquely correct.

Why not just improve the sequence-verification side with ML instead of imaging? One could aim machine learning at the mother-child viral sequence comparison that currently requires manual phylogenetics Rare Detection of Occult Hepatitis B Virus Infection in Children of Mothers with Positive Hepatitis B Surface Antigen. That is worth doing but it is a different problem: it refines the outcome measurement after birth, whereas this proposal targets risk before birth, when intervention (antenatal antiviral escalation) is still possible. In fact we reuse the sequence-homology standard as our endpoint verifier rather than replacing it, because its value against contamination false positives is proven.

Decisive Checks

Four results would decide the proposal. First, the canvas-width experiment: local-backbone Dice declining with growing simulated field of view while the attention variant holds flat, at equal parameter count. Second, out-of-distribution recovery on the real benchmark: posterior-placenta Dice materially above the 0.45 multi-task reference point, with the one-versus-two repetition ablation showing the gain arrives specifically when the receptive field becomes complete. Third, on the clinical cohort, a nested-model comparison in which the image branch adds AUC and net benefit over the serology-only gradient-boosting baseline, with endpoints verified by mother-child sequence homology so that contamination cannot inflate apparent transmission. Fourth, a deployment check: inference within the memory and compute envelope implied by the criss-cross reductions, on hardware representative of the low-resource settings where the epidemiology is worst. No HBV-positive maternal placental ultrasound dataset exists on Lacuna; the honest cost of this proposal is that stage two requires collecting one, and the architectural stage is designed to stand on its own even if the clinical cohort answers the fusion question in the negative.